CHAMPAIGN, Ill. — Online retail platforms are increasingly becoming a dumping ground for low-quality products — which, over time, only serves to tarnish the retailer’s reputation and dent sales. But new research from a team of business scholars at University of Illinois Urbana-Champaign and Penn State points to a new model that would allow e-commerce platforms to proactively identify potentially dubious products for removal.
The findings suggest that a new system architecture — dubbed the “Two-stage classification model” — excels in predicting a sequence of future product ratings for consumer goods and classifying them as problematic or not, said Anton Ivanov, a professor of business administration and the Deloitte Foundation Center for Business Analytics Scholar at the Gies College of Business.
“We developed a system framework that can forecast future ratings of a poor-quality product that is losing support from online shoppers,” he said. “Instead of waiting to take reactive action when a product is already losing altitude and falling out of favor with consumers, our tool is able to look over the horizon and automatically detect where those ratings are trending.
“Ultimately, our model will be an important tool for e-commerce platform managers, since perceived product quality is an important dimension of brand management, and a brand’s reputation directly influences whether consumers purchase a product from an online store.”
Ivanov’s co-authors are Abhijeet Ghoshal, a professor of business administration and the Office of Risk Management and Insurance Research Faculty Scholar at Illinois, and Akhil Kumar, a professor of supply chain and information systems at Pennsylvania State University.
As more consumers shift their shopping from brick-and-mortar stores to e-commerce sites, online reviews and ratings play an increasingly important role for online platforms, whose ultimate aim is to create an “ecosystem of products” sold by tens of thousands of suppliers from all over the world, Ivanov said.
“It’s in the best interests of the platform to ensure that poor-quality products are quickly weeded out and that action is taken against suppliers who consistently supply low-quality products,” he said.
On the other hand, if the platform fails to aggressively police its sellers and their products, its reputation may be sullied, and customers will be skeptical about buying from the platform, according to Ivanov.
“It’s generally hard to monitor and track the quality of products from third-party vendors on an ongoing basis, although platforms do act reactively when too many customers complain,” Ivanov said. “Taking proactive measures would have a huge positive impact for online retailers. In doing so, the platform can maintain its reputation, and customers can feel confident about the quality of the products on offer.”
For their model, the researchers used historical data from an e-commerce dataset — nearly 800,000 observations from more than 2,800 electronics products — to extract useful information from customer ratings and textual reviews. The data was fed into a state-of-the-art deep learning sequence model that first forecasts future ratings and then automatically analyzes the predicted trends. The results show that the proposed two-stage classification approach outperformed a range of alternate single-stage benchmarks, according to the paper.
“We employed a two-stage classification process, which offered a more stable and robust performance than conventional single-stage methods,” Ivanov said. “Our two-stage system makes a rating forecast and then uses the trend of the predicted ratings to classify a product. Ultimately, we found that it fared better than a single-stage benchmark, meaning that these margins of improvements may translate into more products being correctly identified as ‘out of favor’ by our system versus a more simplistic system that only considers previous numerical ratings.”
The implications of the research point to clear benefits for e-commerce platform brands and managers.
“They can benefit considerably from applying this framework to proactively generate a list of potentially dubious products that can be further screened manually to decide what products to eliminate,” Ivanov said. “They also can tweak certain parameters and thresholds in the framework to suit their own needs. Managers can better understand the qualitative factors that lead to products being labeled as ‘out of favor’ by analyzing the associated topics and underlying consumer sentiments.
“Ultimately, our model will help managers of online platforms identify products that are likely to perform poorly in the near future in terms of perceived quality.”
The paper was published in the journal Production and Operations Management.